Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning